Changes in Monetary Regimes and the Identification of Monetary Policy Shocks: Narrative Evidence from Canada

Transcription

1 Changes in Monetary Regimes and the Identification of Monetary Policy Shocks: Narrative Evidence from Canada Julien Champagne Bank of Canada Rodrigo Sekkel Bank of Canada December 23, 217 Abstract We use narrative evidence along with a novel database of real-time data and forecasts from the Bank of Canada s staff economic projections from 1974 to 215 to construct a new measure of monetary policy shocks and estimate the effects of monetary policy in Canada. We show that it is crucial to take into account the break in the conduct of monetary policy caused by the announcement of inflation targeting in 1991 when estimating the effects of monetary policy. For instance, we find that a 1-basis-point increase in our new shock series leads to a 1. per cent peak decrease in real GDP and a.5 per cent fall in the price level, while not accounting for the break leads to a very persistent decrease in real GDP and a price puzzle. Albeit the change in monetary policy regime, we find that the effects of monetary policy have not changed much before and after IT. Finally, we compare our results with narrative evidence for the U.S. and the U.K. and find slightly stronger effect of monetary policy on output and significantly smaller effects on the price level in Canada. Keywords: monetary policy, narrative identification, real-time data, real-time forecasts, business cycles. JEL classification: E31, E32, E43, E52, E58 Contact information: Julien Champagne (julienchampagne@bankofcanada.ca), and Rodrigo Sekkel (rsekkel@bankofcanada.ca). We thank James Cloyne, Oleksiy Kryvtsov, Yuriy Gorodnichenko, John Murray, Larry Schembri, Greg Bauer, Michael Weber, Wataru Miyamoto, Sami Alpanda, Jonas Arias, Thorsten Drautzburg as well as seminar participants at the Federal Reserve Bank of Philadelphia, the CEBRA sessions at the 217 WEAI (San Diego), the 217 conference of the International Association for Applied Econometrics (Sapporo, Japan) and the 217 Bank of Canada Fellowship Exchange conference for useful conversations and comments. We also thank James Cloyne for sharing his Matlab codes, as well as Guillaume Poulin-Bellisle for outstanding research assistance and Hope Pioro and Laurence Savoie for help in retrieving part of the data used in this paper. The views expressed in this paper are solely those of the authors. No responsibility for them should be attributed to the Bank of Canada. 1

2 1 Introduction The identification of monetary policy shocks has generated a vast literature in empirical macroeconomics. Much of this literature uses vector autoregressions (VARs), identified with different approaches, and finds that the effects of monetary policy for the U.S. are relatively modest, with peak decline estimates ranging between.3 and 1 per cent for output following a 1-basis-point monetary innovation. The estimates for the price level range between a positive response (coined the price puzzle by Eichenbaum (1992) and Sims (1992)), to slightly negative, depending on the identification strategy used (e.g., Leeper et al. (1996), Christiano et al. (1996, 1999), Bernanke and Mihov (1998), Bernanke et al. (25), Uhlig (25)). However, Romer and Romer (24, R&R henceforth) find much larger effects of U.S. monetary policy shocks using narrative methods. This strategy uses historical records to construct a series of intended interest rate changes at meetings of the Federal Open Market Committee (FOMC) and then isolate the innovations to these interest rate changes that are orthogonal to the Federal Reserve s information set. A 1- basis-point innovation from the R&R measure translates into output and price level peak declines larger than 4 per cent. Coibion (212) reconciles the differences between R&R and the previous VAR studies and suggests that the effects of monetary policy are more likely to be medium sized (i.e., about -2 per cent for output and prices). Cloyne and Hürtgen (216) extend the narrative approach to the U.K. and find effects of monetary policy that are in line with those of Coibion (212) for the U.S. 1 Another body of literature in monetary economics has documented important shifts in the conduct of monetary policy in the U.S. and abroad. Clarida et al. (2) estimate monetary policy rules for the U.S. and find substantial differences before and after Volcker s appointment, with monetary policy becoming more responsive to expected inflation in the later period. Within a time-varying VAR with stochastic volatility, Primiceri (25) also shows that the systematic component of monetary policy has responded more aggressively to inflation and unemployment. Boivin and Giannoni (26) estimate VARs over the preand post98 periods, and show evidence of a reduced effect of monetary policy shocks in the latter period. Using a small structural model, they show that by responding more 1 For an exhaustive integration of this long literature, see the recent survey by Ramey (216). 2

3 strongly to inflation expectations, monetary policy has been more effective in stabilizing the economy in the post98 period. Focusing on small open economies, Alstadheim et al. (213) show that the central banks of the U.K., Sweden and Canada have shifted away from responding to exchange rate movements around the 199s, when these countries implemented inflation-targeting (IT) regimes. Our paper builds on these two strands of literature and provides a new measure of monetary policy shocks for Canada. We use this measure to bring new evidence on the macroeconomic effects of monetary policy in a small open economy and highlight the importance of changes in the conduct of systematic monetary policy. As in R&R, our analysis proceeds in two stages. The first stage identifies the exogenous component of monetary policy. To do this, we use historical documents to construct a series of intended changes in the target policy interest rate along with a novel database of real-time data and forecasts assembled from the Bank of Canada s staff economic projections from 1974 to 215. These real-time data and forecasts are used to isolate the innovations to the intended policy changes that are orthogonal to the policy-makers information set. We then proceed to the second stage, where we estimate a monthly monetary VAR that includes our new shocks measure as the relevant policy rate. Following a 1-basis-point monetary policy shock, real monthly gross domestic product (GDP) has a peak decline of 1. per cent about 18 to 24 months after the shock, ending close to zero after four years, while the consumer price index (CPI) response is weaker and takes longer to materialize, falling by about.5 per cent after four years. We show that it is crucial for these results to depart from R&R in two important ways: (i) by controlling for U.S. interest rates as well as the USD/CAD exchange rate in the policy-makers information set, and (ii) by accounting for the structural break in the conduct of monetary policy caused by the announcement of IT in Canada has had a continuing floating exchange rate regime since 197 and exchange rate movements were an important factor in monetary policy decisions, as documented by Fortin (1979), Courchene (1981), Lubik and Schorfheide (27) and Alstadheim et al. (213), among others. Additionally, the announcement of the IT regime in 1991 prompted a sharp shift in the conduct of monetary policy. We show that while the exchange rate and U.S. interest rates were the main determinants of changes in the target policy rate in the 197s and 198s, 3

4 GDP and inflation forecasts have become the key factors since the introduction of the IT regime. Furthermore, we show that not accounting for this change leads to monetary policy shocks that imply a prize puzzle and show clear signs of endogeneity. Given the break in the conduct of monetary policy in Canada, we investigate whether the effects of monetary policy have changed after the introduction of IT. We use our new measure of shocks and estimate our baseline VAR separately for both our subsamples (pre- IT and IT periods). We find that although real GDP reacts modestly more and the price level slightly less during the IT period, the responses are qualitatively similar. These results sharply contrast with those of Barakchian and Crowe (213) who find that since 1988, a contractionary monetary policy shock (from commonly-used VARs and the R&R shocks) raises output and inflation in the U.S. We then compare the effects of monetary policy in Canada with narrative evidence from the U.S. (R&R shocks updated to 27) and the U.K. (Cloyne and Hürtgen (216) shocks). When comparing our results, we also control for a break in the monetary policy reaction function for those countries. 2 We find that monetary policy shocks in Canada generate output effects that are in line with those of the U.S. and somewhat stronger than in the U.K.; however, the price level response is weaker in Canada than in those countries. This exercise further shows that, on the one hand, accounting for a break in the monetary policy reaction function matters importantly when identifying shocks for the U.S., consistent with Barakchian and Crowe (213) and Ramey (216). On the other hand, as argued by Cloyne and Hürtgen (216), accounting for the introduction of IT in the U.K. is not as important when estimating the effects of monetary policy, in contrast with our evidence from Canada. We further examine the robustness of our main results to many alternative specifications and provide additional results: for instance, (i) using Jordà (25) s local projections method, we test the robustness of our results and also assess the responses of open economy variables such as the USD/CAD exchange rate, imports and exports. Following a 1-basis-point shock from our new measure, we find that the Canadian dollar appreciates (relative to the U.S. dollar) by about 1. per cent for a year and then depreciates slowly, 2 The conduct of monetary policy in the U.S. and the U.K. has also undergone significant changes. As just mentioned, many papers argue that U.S. monetary policy has become more responsive to future expected inflation since the Volcker disinflation. Similarly, the U.K. also introduced an IT regime in 1993, and granted the Bank of England formal autonomy to pursue these targets in

5 ending 2 per cent lower after four years, leading to a peak decline of 4. per cent in imports. Following the decline in economic activity, exports decrease by about 3.5 per cent despite the depreciation of the currency; (ii) we estimate a standard VAR à la Christiano et al. (1996, 1999) with the target policy rate instead of our shocks series and find a large price puzzle; (iii) we show the robustness of our baseline VAR results with different measures of output and price level; and finally, (iv) we show that our results are robust to many alternative first-stage specifications. The remainder of the paper is organized as follows. Section 2 provides a brief historical discussion of the Canadian monetary policy framework, details the derivation of the new measure of monetary policy shocks for Canada and provides some analysis. Section 3 presents our baseline results for the effects of monetary policy and compares them with narrative evidence from the U.S. and the U.K. Additional results and robustness exercises are presented in Section 4 and Section 5 concludes. 2 Derivation of a new measure of monetary policy shocks for Canada In this section, we first provide a short historical discussion of the Canadian monetary policy framework and then describe how we adapt R&R s identification strategy to these institutional details. Next, we specify our first-stage regression and provide a description of our database construct of real-time data and staff forecasts. Finally, we summarize the first-stage estimation results, present the new measure of monetary policy shocks and provide some analysis. 2.1 Monetary Policy in Canada: Background The Bank of Canada was established in 1935, in the middle of the Great Depression, in response of a severe deterioration in public confidence in the behavior of markets and the financial system. Since then, it has been operating under the Bank of Canada Act, which is, in essence, similar to the broad mandate given to the U.S. Federal Reserve under 5

6 U.S. legislation. 3 Although the Bank s mandate remained broadly unchanged since its beginnings, the monetary policy framework and the targets for inflation control changed significantly. For example, in the early 197s, the Bank s efforts to curb inflation in the face of international inflation pressures proved to be insufficient; as a result, in 1975 the Bank followed other central banks and began targeting monetary aggregates (M1). quickly proved to be an ineffective strategy and was dropped in It The Bank then entered a period (up to 1991) in which there was no clear monetary policy target; there was rather only a desire to reduce inflation and to approach price stability. After the disinflation, there was no further reduction in inflation, and without an explicit target for monetary policy, the Bank had difficulty explaining its policy actions. Discussing that turbulent period, Thiessen (2) states: Any efforts to keep interest rates low and to support economic activity were often misinterpreted as signs that the Bank was, in fact, pursuing an inflationary policy and would only make matters worse. As a result, the Bank found itself having to follow the U.S. lead on interest rates and resist downward movements in the exchange rate for tactical reasons. Failure to defend the Canadian dollar would produce even larger increases in domestic interest rates and inflict more serious damage on the economy. Indeed, before the 199s the Bank seemed to follow U.S. rates closely and respond to changes in the value of the Canadian dollar relative to its U.S. counterpart. For example, Fortin (1979), Courchene (1981), Kuszczak and Murray (1986) and various press releases (e.g., Bank of Canada (1992, 1993)) all point to the Bank reacting to movements in U.S. rates and/or changes in the USD/CAD exchange rate at different points in time. 4 Given the relatively high rates of inflation throughout the 198s, the economic boom and the introduction of the goods and services tax by the federal government in 1991, policymakers were afraid that inflation would again escalate. 5 It was against this backdrop that 3 The preamble to the act notes that the central bank is established to regulate credit and currency in the best interests of the economic life of the nation, to control and protect the external value of the national monetary unit and to mitigate by its influence fluctuations in the general level of production, trade, prices and employment, so far as may be possible within the scope of monetary action, and generally to promote the economic and financial welfare of Canada. 4 Our results in Section 2.4 below show that the Bank was reacting aggressively to exchange rate movements prior to the early 199s. 5 This period was marked by different developments that clouded the economy with uncertainty. For example, after five years of relative stability at around 4 per cent, inflation took off in 1988; the Bank Rate was very high, coupled with the historical high of the Canadian-U.S. 9-day treasury bill spreads; and the USD/CAD exchange rate was very volatile. There was also political uncertainty around the unravelling of 6

7 the Bank of Canada and the federal government agreed on targets for inflation reduction in February The first inflation targets were set at 3 per cent by the end of 1992, 2.5 per cent by mid994 and 2 per cent by the end of Inflation fell more rapidly than expected; by the end of 1992 it had already converged to 2 per cent. This target has become the official operating objective for monetary policy, although inflation is allowed to vary within a range of 1 to 3 per cent. 6 Although the framework around which the Bank conducted monetary policy evolved substantially over the past four decades, the Bank s main tool to conduct policy has always been an explicit interest rate to influence the overnight market (Courchene (1979), Fettig (1994)). 7 Since February 1996, the Bank uses the Target for the Overnight Rate to indicate its intention for the overnight market, and this rate is set as the mid-point within a 5 basis points operating band. The Bank Rate is set as the upper limit of this band. Note that the operating band for the overnight rate was introduced in April 1994, nearly two years before the introduction of the target rate. The introduction of the operating band was seen as a first step to provide more transparency to monetary policy (Lundrigan and Toll (1998)). 8 Prior to the operating band for the overnight rate, the Bank Rate has always been the key interest rate at which the Bank lends to chartered banks. Finally, since December 2 the Bank takes monetary policy decisions at fixed announcement dates (FAD), eight times per year. These dates are determined and announced to the public about one year and a half before the first FAD of the year. the Meech Lake Accord in Since the introduction of the inflation-targeting regime, the Bank s policy-makers have relied heavily on staff economic forecasts and other current economic indicators to determine the appropriate stance of monetary policy. We provide evidence in Section 2.4 below. 7 Even when the Bank emphasized targeting the money supply (during the period), it achieved the targeted growth rate of money by setting short-term interest rates at levels that would bring the monetary conditions consistent with the Bank s views of inflationary pressures, and not through changes in the reserve base, which would have influenced the money supply (e.g., Fortin (1979), Sparks (1979), Racette and Raynauld (1994) and Longworth (23)). 8 These rates are published on the Bank of Canada website. See Appendix A.2 for more details on the key interest rates at the Bank of Canada. 7

8 2.2 Identification The intended monetary policy variable (e.g., the target policy rate) is a combination of a systematic component f(ω t ), where the function (f) captures the policy-makers reaction function to an information set (Ω t ) and an exogenous component (ɛ t ) that reflects unexpected changes in monetary policy. R&R s strategy to identify ɛ t the narrative approach has two key steps: first, they derive a series of intended changes in the federal funds rate (FFR) around FOMC meetings, and second, they control for the Federal Reserve staff real-time forecasts (Greenbooks) to create a measure of intended FFR changes that is orthogonal to information about past, current and future economic developments. 9 This approach is appealing because it goes a long way in overcoming many econometric challenges faced when estimating the effects of monetary policy, such as the simultaneous determination of interest rates and macroeconomic variables, 1 behavior and real-time nature of monetary policy decisions. 11 the forward-looking However, R&R s approach implicitly assumes that the monetary policy reaction function does not change throughout the sample they study. This is at odds with some prominent papers such as Clarida et al. (2), Primiceri (25) and Boivin and Giannoni (26), which show a significant change in the systematic component of U.S. monetary policy around the early 198s, where policy began reacting more aggressively to expected future inflation. We apply the R&R approach to identify monetary policy shocks (ɛ t ) for Canada. 12 One of the major contributions of R&R was to use the minutes of the FOMC meetings to construct the series of intended changes in the FFR, as the Federal Reserve did not always 9 R&R s strategy has become a popular method to derive measures of exogenous shocks. Apart for monetary policy shocks, the narrative approach has also been employed to identify fiscal policy shocks; e.g. Romer and Romer (21) and Ramey (211) for the U.S. and Cloyne (213) for the U.K. For a recent application to government asset purchases, see Fieldhouse and Mertens (217) and Fieldhouse et al. (217). 1 Many studies have tackled the simultaneity problem using a timing restriction in VARs (e.g., the recursive assumption of Christiano et al. (1996, 1999)); others have used factor-augmented VARs (e.g., Bernanke et al. (25)) to exploit a larger set of data as a better proxy for the policy-makers information set, or used sign restrictions (e.g., Faust (1998), Uhlig (25)) as a way to identify the shocks ɛ t. 11 Central banks devote a lot of resources monitoring the economy and forecasting in real-time the future path of macroeconomic variables (e.g., see Romer and Romer (2) and Croushore and Stark (21) for the Federal Reserve (U.S.); Duguay and Poloz (1994), Poloz et al. (1994) and Dorich et al. (213) for the Bank of Canada). Moreover, there can be stark differences in using ex-post revised vs. real-time data when estimating monetary policy reaction functions (e.g., Orphanides (21, 23) or Molodtsova et al. (28)). 12 We use the term narrative because we use various historical documents to (i) construct the intended policy target; (ii) construct our database of real-time data and staff forecasts; and (iii) justify our estimation of equation (1) below. 8

9 target it explicitly. 13 As mentioned above, a handy feature of the Canadian monetary policy framework is that the Bank of Canada has always used an explicit target interest rate. Specifically, in the first step we construct our policy rate series using the Target for the Overnight Rate from February 1996 onward, the operating band for the overnight rate between April 1994 and February 1996, and the Bank Rate between 1973 and April This framework is very similar to the U.K., where the policy rate (also called the Bank Rate) has always been the relevant policy target. 15 The next step in the derivation of our monetary policy shock series is to purge the intended policy rate from the systematic component f(ω t ) of monetary policy. The specific regression equation we estimate is: i m =α + β 1 i t d14 + ρ h u t h + γ j ŷ f m,j + δ j π f m,j h=1 j= 1 j= θ j (ŷ f m,j ŷf m 1,j ) + φ j (π f m,j πf m 1,j ) j= 1 j= 1 (1) + β 2 F F R t d14 + β 3 ER t d14 + β 4 F F R m m 1 β 5 ER m m 1 + ɛ m, where the dependent variable ( i m ), the change in the intended policy rate, is measured at a meeting-by-meeting frequency, as indicated by the subscript m. The subscript j denotes the quarter of the real-time data or forecast relative to the meeting date, while subscripts t h and t d14 refer to information from the previous months and two weeks relative to the meeting date, respectively (and not to information from a previous meeting). Specifically, we regress the change in the policy target rate ( i m ) between two meetings on the oneand two-quarter-ahead forecasts of real output growth (ŷ f m,j ) and inflation (πf m,j ), as well as the nowcast and the real-time one-quarter lag. 16 We also include the revisions to the forecasts relative to the previous round of forecasts (e.g., ŷ f m,j ŷf m 1,j ), since both the level 13 The Volcker s disinflation period is an example of when the Federal Reserve was not explicitly targeting the FFR. 14 See Appendix A.2 for more details on the construction of our intended policy rate series. 15 Also note that the Bank Rate, the operating band for the overnight rate, or the Target for the Overnight Rate are not market rates but policy rates announced by the Bank of Canada. It is similar to the target for the FFR for the U.S., and to the U.K. Bank Rate used by Cloyne and Hürtgen (216). 16 Note that we test the robustness of our results to adding a second lag (t 2) of real-time GDP growth and inflation (see Section 4.4) and to adding a third quarter (t + 3) of forecasts (see Appendix D.7). We find that our results are unaffected by these alternative first-stage specifications. 9

10 and change in the forecasts can be important determinants of the Bank s behavior. To control for economic developments between meetings, we include the intended policy rate two weeks before the meeting and the (real-time) rates of unemployment for the previous three months. 17 The third line of equation (1) departs importantly from R&R as we also control for the levels and changes of the U.S. FFR (F F R t d14 ) and the logarithm of the USD/CAD nominal exchange rate (ER t d14 ) two weeks before the meeting. 18 Canada is a small open economy with close ties to the U.S., and these variables are included to capture any tendency for the Bank to react to interest rate movements in the U.S. as well as the changes in the value of the Canadian dollar relative to its U.S. counterpart. As mentioned above, the Bank reacted to movements in U.S. rates and/or changes in the USD/CAD exchange rate at different points in time, especially before the introduction of the IT regime. 19 We make another important departure from R&R and break the estimation of equation (1) into two sub-samples of meetings: the first sub-sample includes all those meetings preceding the inflation targets (i.e., ) and the second sub-sample regroups all meetings afterward (1992 on). 2 The reason for doing this is three-fold: first, there is strong evidence suggesting a change in the Bank s reaction function. For example, Rowe and Yetman (22) argue that there was a major change in the Bank s objectives near the time formal inflation targets were announced. As explained in the previous subsection, the Bank followed U.S. interest rates and exchange rates developments very closely before the IT period, while since the beginning of IT, the Bank has been using economic forecasts and other current economic indicators more thoroughly to assess the necessary stance of 17 As mentioned by R&R, using forecasts to identify monetary policy shocks has a further advantage since they summarize a wider range of macroeconomic information as well as the anticipated movements in the economy. This approach thus allows us to identify shocks without including a large set of variables as in the FAVAR approach, for example. 18 In Section 4 below, we show that our results are robust to using a seven-day lag instead of 14 days. 19 Note that we control for the level and change of the U.S. FFR and the USD/CAD prior to a meeting m and not for the forecasts. The reason is two-fold: first, the narrative records point to the Bank reacting to changes in U.S. FFR and the USD/CAD nominal exchange rate, especially during the 198s, and not responding to anticipated future movements in those rates. Second, we do not have a consistent set of forecasts for these variables. 2 Because IT was announced in 1991 but the first inflation target was set for 1992, we decide to break the first-stage estimation between 1991 and 1992; consequently, our first sub-sample covers and our second covers 1992 to 215. Note that breaking the estimation right after the announcement (between February and March 1991) does not alter our results. 1

11 monetary policy (e.g., Montador (1995), Duguay and Poloz (1994), Macklem (22), Dorich et al. (213) and various Monetary Policy Reports (MPRs)). 21 Second, we will show in Section 2.4 that our regression estimates of equation (1) strongly support this change in the reaction function, as the response of the intended policy rate ( i m ) to changes in the USD/CAD exchange rate and U.S. FFR and to real GDP growth and inflation forecasts changes drastically in the first vs. the second sub-sample. Third, as we will show in Section 3, the effects of monetary policy are markedly different when one does not break the first-stage estimation to account for the IT period. 22 Lastly, we take the estimated residuals from equation (1) for each sub-sample and splice them together to create our new meeting-by-meeting series of monetary policy innovations. This new series, which can be interpreted as exogenous changes that are not taken in response to information about current and future economic developments, will then be converted into a monthly basis and used in the second stage (Section 3) to quantify the effects of monetary policy on macroeconomic variables. 2.3 Data construction When constructing our dataset for the first-stage regression (equation 1), we need to match the variables forming the Bank s information set (Ω t ) with the intended policy rate variable. As noted above, our intended policy rate is constructed using the Bank Rate from 1973 up to March 1994, the operating band for the overnight rate between April 1994 and February 1996, and the Target for the Overnight Rate afterward. We use the changes in this series as the left-hand side variable in equation (1). To build the information set (Ω t ), we first use a novel database of real-time data and forecasts for real GDP and inflation constructed from the Bank of Canada s staff economic projections. Bank of Canada staff produce four exhaustive projections each year, following the release of the quarterly national income and expenditure accounts, which are generally 21 For example, Courchene (1981) has a whole chapter titled 198: Riding the U.S. roller coaster, arguing that Canada was simply a side passenger in the U.S. monetary policy roller coaster of the early 198s. Also, Howitt (1986) states that the Bank s policy after 1979 consisted of resisting any long-term depreciation of the Canadian dollar and short-term fluctuations in that rate caused by temporary changes in U.S. interest rates. 22 For sake of comparison and completeness, in Section 4.2 we use the same specification as in R&R to identify the shocks and compare their implications with those from our new measure of shocks. 11

12 carried out around the end of March, June, September, and December. 23 These staff projections contain quarterly forecasts as well as historical (real-time) data of numerous macroeconomic aggregates. They are a material part of the analysis presented to the Governing Council every quarter in the weeks leading up to the publication of the Bank s Monetary Policy Report. 24 The quarterly staff projections are analogous to the Greenbook forecasts prepared by the Federal Reserve Board staff; they are judgmental in the sense that the forecasts are based on different sources of information and economic models. Second, we construct the (real-time) unemployment rate series from digitized Statistics Canada archives ( ) and hard copies of Bank of Canada Reviews for Third, our series for the USD/CAD exchange rate is taken from Statistics Canada, and the U.S. FFR is taken from the Federal Reserve Board website. Appendix A.1 presents further details on the data series used in the first-stage regression. This is a large dataset that we hope will be useful for future research. 25 Note that the relevant inflation index varied over our sample. Up to March 198 inclusively, we use total CPI inflation, while from April 198 onward, we use core CPI inflation. 26 The output measure forecasted by staff also varied: prior to 1982, staff prepared forecasts for real gross national expenditures, while since 1982, it has been forecasting real GDP growth. The earliest vintage of projection data where we have both output and inflation forecasts is 1974:Q1. Before matching each monetary policy decision ( meetings, for short) with the relevant set of real-time data and forecasts, we need to define what exactly we mean by meeting. As mentioned above, since December 2 the Bank announces monetary policy decisions 23 Note that for many years from 25 on, Bank of Canada staff produced eight projections per year: the four just mentioned, and four updates prepared for the FADs in-between the national accounts data releases. When available, we use these additional forecasts. 24 See Macklem (22) for details about the information and analysis presented by the staff to the Governing Council. We highlight the fact that these are staff estimates and thus may not be the same estimates provided in the MPR, the Bank of Canada publication containing detailed economic analysis and economic outlook and representing the view of the Governing Council, available every quarter since For example, Champagne et al. (217) use the same projections dataset to study the real-time properties of the Bank of Canada s staff output gap estimates. 26 Since May 21, core CPI inflation in Canada has been known as CPIX, which excludes the eight most volatile components of the CPI and adjusts the remaining components for the effects of changes in indirect taxes. Before May 21, core inflation was defined as CPI excluding food, energy, and the effects of changes in indirect taxes. 12

13 eight times per year on predetermined dates (FAD). Consequently, these FAD become our meetings from 2:12 to 215:1 (since the latest policy decision in our sample is October 21, 215). Before 2:12, the definition of meeting is less trivial as we cannot know with certainty when a meeting happened. Following Cloyne and Hürtgen (216), we define a meeting as all dates where a change in our intended policy rate occurred; moreover, when a new set of forecasts is released but there is no change in our intended policy rate, we do not treat the forecast release as a decision itself because we cannot be sure these are genuine monetary policy decisions. 27 Next, we need to assign the relevant projection data to each meeting since the regression in equation (1) is conducted on a meeting-by-meeting frequency. We are facing the complication that we do not have a new projection for every monetary policy decision because there are more meetings than forecast releases. For all those meetings without a new projection, we assign the latest available set of forecasts. 28 Note that since the introduction of the fixed announcement dates in December 2, this has been less of an issue because a new set of forecasts is prepared for most of the announcement dates such that matching a new projection with a meeting is straightforward. 29 Before this, we must be careful when assigning a set of forecasts to a given meeting in order to avoid endogeneity of forecasts to the policy change. Therefore, we use the projection prepared at the end of the quarter preceding the meeting date, ensuring that the forecasts do not include the effects of the subsequent policy change. Table (1) provides examples of the data assignment to meetings. The first column lists the variables of interest (i.e., regressors) from equation (1), while the other columns show the data source and the time period forecasted (or the backdata, when using lagged data) for different meeting dates (shown in the top row). The second block of columns, corresponding to the April 17 and May 22, 198 meetings, 27 In Section 4.4, we test the robustness of the definition by adding meetings with zero policy rate change in two different ways: (i) we add all those dates where we have a new forecast but no change in the intended policy rate and (ii) we add those dates corresponding to two days after an FOMC meeting and where we do not have a meeting already defined in the week. We find that our results remain broadly robust to adding these additional observations. 28 Note that in equation (1) we control for developments between a given meeting and the last available projection by including monthly lags of the unemployment rate and the interest rate two weeks before the meeting. 29 For those announcement dates without a new projection, we keep assigning the latest projection data available. 13

14 Table 1: Assigning forecasts and economic variables to monetary policy decisions Meeting dates [current quarter] 12/22/1976 [1976:Q4]... 4/17/198 [198:Q2] 5/22/198 [198:Q2]... 7/17/21 [21:Q3] Variables Source Forecast / Data Source Forecast / Data Source Forecast / Data Source Forecast / Data ŷ f Sep-76 March-8 March-8 June m,t :Q :Q1 198:Q1... proj proj proj proj 21:Q2 ŷ f Sep-76 March-8 March-8 June m,t 1976:Q :Q2 198:Q2... proj proj proj proj 21:Q3 ŷ f Sep-76 March-8 March-8 June m,t :Q :Q3 198:Q3... proj proj proj proj 21:Q4 ŷ f Sep-76 March-8 March-8 June m,t :Q :Q4 198:Q4... proj proj proj proj 22:Q1 π f Sep-76 March-8 March-8 June m,t :Q :Q1 198:Q1... proj proj proj proj 21:Q u t 1 Real-time Real-time Real-time Real-time Nov Mar-8 Apr-8... data data data data Jun u t 2 Real-time Real-time Real-time Real-time Oct Feb-8 Mar-8... data data data data May ER t d14 Real-time Real-time Real-time Real-time 12/8/ /3/8 5/8/8... data data data data 7/3/1 F F R t d14 Real-time Real-time Real-time Real-time 12/9/ /3/8 5/8/8... data data data data 7/3/ Notes: Assignment of forecasts and lagged real-time data to different meeting dates. The upper row corresponds to the exact date of the meeting, with (in brackets) the corresponding year and quarter in which the given meeting is happening. The Source column refers to the specific staff projection database from which the real GDP growth and inflation real-time and forecasts data were prepared, while the lagged (real-time) unemployment rate, exchange rate, FFR and the policy rate are taken from Statistics Canada or Bank of Canada archives. The Forecast/Data column shows the year and quarter of the forecasts estimates (for real GDP growth and inflation), while it shows the specific date (month or day) for the other lagged real-time variables. provides a case where we assign the same set of forecasts to two consecutive meetings. Because these two meetings happen in 198:Q2, we use the projection data from March 198 to get the values for real GDP growth (ŷ f m,t+j ) and inflation (πf m,t+j ). Moreover, we assign the relevant lags of unemployment rates, exchange rates and U.S. FFR available at the time of the meeting from our real-time data set. 3 Lastly, we use all changes in our intended policy rate, apart from those meetings occurring within the same four weeks during the March 198 to May 1994 period (when meetings were occurring at a high frequency). 31 Overall, our data assignment allows us to have a 3 Note that the change in the USD/CAD exchange rate and in the U.S. FFR (i.e., ER m m 1, F F R m m 1 in equation (1)) between these two consecutive meetings (May 22 and April 17, 198) are computed as follows: (ER 22May198,t d14 ER 17Apr198,t d14 ) and (F F R 22May198,t d14 F RR 17Apr198,t d14 ), respectively. 31 During those years the Bank Rate was changing more often than in the rest of the sample (see Appendix A for details). We test the robustness of our results in Section 4.4 to dropping changes within two weeks instead of four weeks for the specified period and find that both specifications yield similar results. 14

15 sample of meetings matched with projections data from 1974:M4 to 215:M1, containing a fair number of observations (337 meetings). 2.4 Estimation results We now use our dataset of changes in the intended policy rate, carefully matched (by meeting) with the staff forecasts and real-time data to estimate the regression equation (1). As mentioned earlier, our preferred strategy critically departs from R&R as we estimate equation (1) for the pre992 and 1992-onward periods separately, such that our estimates account for the change in the monetary policy reaction function observed around the introduction of IT. To show the importance of breaking our estimation into two parts, we also estimate equation (1) using the full sample (all 337 meetings). Table (2) reports the results of these estimations. Examining the full-sample (no-break) estimates (left column) suggests that monetary policy has been conducted in an acyclical fashion over the past 4 years. For example, coefficients on real GDP growth level sum up to.27, while they sum to -.24 for the revision to the forecast. Coefficients on inflation forecast levels and revisions to inflation forecasts are very low, at.4 and -.2, respectively. Thus, a 1 percentage point increase in inflation from one forecast release to the next is associated with a mild increase in the policy rate of.2. Finally, a 1 percentage point fall in the unemployment rate translates into a small.6 percentage point increase in the policy rate. Monetary policy is also positively related to changes in the U.S. FFR and negatively related to movements in the USD/CAD exchange rate. This acyclical behavior of monetary policy, along with the responses to U.S. interest rates and exchange rate movements, hide considerable heterogeneity once one breaks the first-stage estimation into two sub-samples. The second and third columns (the IT-break estimation ) show the regression estimates for the and the sub-periods, respectively. 32 Three results stand out: first, summing coefficients on real GDP and inflation for the pre-it period yields procyclical estimates of -.17 and -.29 percentage points, respectively, and a countercyclical estimate of -.18 for the unemployment rate. A very 32 A Chow test of parameter stability with a break date set to January 1992 yields an F-statistic of 2.26 with an associated p-value of.7. 15

16 Table 2: Determinants of the change in the policy rate Full-sample (no-break) estimation IT-break estimation Pre-IT IT period Variable Coefficient Standard error Coefficient Standard error Coefficient Standard error Constant.166 (.29) (1.57) (.3) Initial Bank Rate -.29 (.3).6 (.8) -.53 (.4) U.S. FFR: level.45 (.3).46 (.7).13 (.3) change.292*** (.7).249*** (.8).164* (.9) US/CAD exchange rate: level.71 (.27).186 (1.69) -.13 (.22) change -3.73*** (1.35) 4.139*** (4.).222 (1.8) Forecasted output growth, Quarters ahead:.223 (.14).511** (.21).13 (.9) (.22) (.27) (.23) (.24).449* (.25).3 (.4) 2.39 (.2).7 (.3).138 (.25) Change in forecasted output growth, Quarters ahead: -.33 (.12) -.26 (.16) -.46 (.1).225 (.19) -.59 (.25).363* (.2) (.2).76 (.22) -.498* (.29) (.24) -.572* (.34).62 (.18) Forecasted inflation, Quarters ahead:.19 (.13).74 (.21) -.8 (.2) (.2) -.52 (.25).73 (.72) 1.38** (.17).185 (.18) 1.79 (1.29) (.13) -.353** (.15) (1.9) Change in forecasted inflation: Quarters ahead: -.84 (.18) (.26) -.93 (.19) -.12 (.24) (.44) 1.123* (.59) (.28).27 (.35).6 (.95) 2.1 (.33).43 (.51).415 (.82) Unemployment rate, : -.333* (.17) -.39* (.23) -.24 (.21) (.22).339 (.34).93 (.18) (.15) (.22).129 (.12) Observations R-squared Notes: Robust standard errors in parentheses; asterisks indicate statistical significance (i.e., (***: p<.1, **: p<.5, *: p<.1). Dependent variable: change in the intended policy rate, constructed as described in the text. Full-sample estimation refers to first-stage regression estimated over the full sample (1974:M4 215:M1). IT-break estimation refers to first-stage regression estimated over two sub-samples separately (i.e., Pre-IT: , and IT period: ). A Chow test of parameter stability with a date break fixed to January 1992 (IT introduction) yields a F-statistic of 2.26 with an associated p-value of.7. 16

17 different picture emerges for the second sub-sample: summing the same coefficients yields estimates (in percentage points) of.29 or real GDP,.68 for inflation and -.2 for the unemployment rate, implying strong countercylical monetary policy since Second, the pre-it behavior of monetary policy can be explained by the response with respect to the U.S. rates and USD/CAD exchange rate: a 1 per cent increase in the U.S. FFR between two meetings translates into a.3 percentage point increase in the policy rate, while a 1 per cent decrease in the value of the exchange rate between two meetings implies a.15 percentage point increase in the policy rate. Third, the response of monetary policy toward these two variables changed dramatically since 1992: the response to changes in the FFR decreased by more than a third (.18) and fell to almost zero with respect to exchange rate movements (.1 percentage point). Since 1992, Canadian monetary policy has been responding fiercely to economic developments related to inflation and real GDP growth and much less to exchange rate movements. Overall, our full-sample estimates for the response of the policy rate to real GDP growth and inflation forecasts are close to zero and far from the procyclical first-stage estimates of R&R for the U.S. and Cloyne and Hürtgen (216) for the U.K., and are not a good representation of the Bank s behavior for more than half of our sample. Looking at our estimation results, we see they are more similar than those of the U.S. and U.K., albeit not identical. Finally, the residual component of equation (1) from the ITbreak estimation (ɛ m ), i.e., the component of policy rate changes that is orthogonal to the policy-makers information set, is our new measure of monetary policy shocks. 2.5 Analyzing the new shock series This residual series corresponds to specific meeting (i.e., policy decision) dates. To use these residuals for economic analysis, they must be converted into a monthly series of monetary policy innovations. To do this, we assign each shock to the month in which the corresponding meeting occurred. If there is no meeting in a given month, we record the shock as zero for that month; if there is more than one meeting within the same month, we sum the shocks. Figure 1 presents our new monthly series of monetary policy shocks, which we denote ɛ t as above. 17

18 Figure 1: New monthly monetary policy shocks series for Canada 4 3 Percentage points Notes: Shaded gray bars represent recessions as determined by the C.D. Howe Institute. As found by R&R for the U.S. Cloyne and Hürtgen (216) for the U.K., our new series for Canada is more volatile in the first half of our sample, up to the end of Note that our series is substantially more volatile in the early 199s than the U.S. and U.K. series; as mentioned earlier in Section 2.1, this is likely due to the uncertain environment the Canadian economy was facing in those years. The decline in the magnitude of the new measure of shocks coincides nicely with the view that there was a regime change in 1992, when the Bank of Canada began targeting inflation explicitly. Three other developments mentioned earlier also made the policy-making process more transparent during this subperiod: the introduction of the operating band around the overnight rate in 1994, which led to the explicit Target for the Overnight Rate in 1996, as well as the introduction of fixed announcements dates in December We can use our new measure of monetary policy shocks to measure the stance of monetary policy: i.e., periods with a sequence of positive innovations are ones in which the 33 As noted by Cloyne and Hürtgen (216), larger shocks in the first part of the sample could also reflect that the level of the Bank Rate was relatively higher than in the second part. 18

19 Bank raised the policy rate more than it would normally have given current and expected economic conditions. Appendix Figure B.1 plots the exogenous Bank Rate path (i.e., our new shock series, cumulated) along with the shocks series estimated over the full sample (no break) and the actual path of the Bank Rate. Interestingly, the shocks series imply a very different path for the exogenous policy rate (dotted blue vs. dashed red lines). For instance, after following a pattern similar to our new measure of shocks between 1974 and 1989, the no-break series diverged markedly afterwards: it implies that monetary policy had been very tight for 1 years up to 1999, where it loosened precipitously afterward. Since 27, this alternative shock measure implies that given current and expected future economic conditions, monetary policy has been by far at its loosest stance over the last 4 years. 2.6 Are the new estimated monetary policy shocks predictable? We follow Coibion (212) and test whether our new monthly measure monetary policy shocks is unpredictable from movements in ex-post revised data. 34 We perform a Granger causality test by regressing our innovations series ɛ t on a large set of lagged macroeconomic variables (x t i ) including two of the most relevant measures of inflation for Canada (CPI, CPIX), two measures of output (real GDP, industrial production), the unemployment rate, commodity price inflation, the change in the Toronto Stock Exchange Index (TSX) and the change in money supply (M2): 35 ɛ t = c + I β i x t i + υ t. (2) i=1 Under the null hypothesis that our shock series ɛ t is not predictable, the β i are jointly equal to zero. Table 3 reports the F-statistics and p-values for the null hypothesis based on estimation of equation (2) for our new measure of monetary policy innovations ( New measure of shocks, right panel) along with the alternative shock series discussed above, 34 We also look whether our new measure is uncorrelated with other structural shocks, such as the U.S. and U.K. monetary policy shocks of R&R and Cloyne and Hürtgen (216) (correlations of.3 and.6, respectively), as well as oil supply shocks (.) from Kilian (29). 35 See Appendix A.3 for data details. Note that we have a monthly series for real GDP in Canada, which we will also use below when quantifying the macroeconomic effects of monetary policy shocks. 19

20 estimated using the full sample ( no-break, left panel). Table 3: Predictability of monetary policy shocks series Full sample (no-break) shocks New measure of shocks I = 3 lags I = 6 lags I = 3 lags I = 6 lags Variable F-stats P-values F-stats P-values F-stats P-values F-stats P-values CPI inflation CPIX inflation Change in real GDP Change in ind. prod Unemployment rate Commodity price inflation Change in nominal USD/CAD Change in TSX Money growth (M2) Notes: The table reports F-statistics and p-values for the null hypothesis that all coefficients (β i ) are equal to zero. The standard errors are corrected for the possible presence of serial correlation and heteroskedasticity using a Newey-West variance-covariance matrix. The Full sample (nobreak) shocks specification refers to first-stage regression estimated over the full sample ( ). New measure of shocks refers to our new series of monetary policy shocks, where the first-stage regression is estimated over two sub-samples separately (i.e., Pre-IT and IT period). Estimation sample from 1974:M1 to 215:M1. All variables are at monthly average frequency. Two things stand out: first, the shock series estimated on the full sample (no-break) shows many low p-values, implying some degree of predictability. For example, this alternative shock series is significantly predictable using real GDP growth, commodity price changes and six lags of the unemployment rate; it also has low p-values when equation (2) is estimated with the CPIX and the change in the TSX. Second, a sharp contrast emerges when one looks at our new measure of shocks (columns 5 to 8): for example, the p-values of real GDP are almost 5 times (three lags) and 1 times (six lags) larger using our new measure and substantially larger for the change in commodity prices and the unemployment rate. Overall, most p-values are very large for the new measure; this lack of predictability and stark contrast with the alternative full-sample shock series is another argument in favor of our new measure of monetary policy shocks and suggests that it is a suitable instrument for identifying the macroeconomic effects of monetary policy in Canada. 2